Author(s): FAITY, G., CHANTEAU, A., MAHE, G., LE FAUCHEUR, A., Institution: UNIV RENNES, M2S (LABORATOIRE MOUVEMENT, SPORT, SANTÉ), F-35000 RENNES, FRANCE, Country: FRANCE, Abstract-ID: 2571

INTRODUCTION:
Physical activity (PA) is associated with major health benefits. Since 2020, the World Health Organization no longer requires PA to be accumulated in bouts of at least 10 minutes. Thus, brief and intermittent episodes now contribute to recommendations (Bull et al., 2020).
However, many wearables still rely on 60-s epoch aggregation due to historical technical constraints and alignment with previous guidelines. Such temporal resolution is poorly suited to capture the highly fragmented nature of daily-life PA (Taoum et al., 2021). Moreover, most validation studies are conducted in controlled settings, limiting ecological validity (Giurgiu et al., 2025). Although newer approaches use shorter epochs, few studies have directly compared these methods under free-living conditions.
This study compared the validity of several devices and analytical methods to detect walking bouts and classify their intensity in free-living conditions using video-recorded direct observation.
METHODS:
Sixty healthy adults (20–80 years, balanced by sex and decade) simultaneously wore six devices for seven days: activPAL (thigh), StepWatch4 (ankle), ActiGraph wGT3X-BT (hip and wrist), Fitbit Charge 5 and Inspire 2 (hip and wrist), and Garmin Vívoactive 4S (wrist). A total of 119 free-living sessions were annotated by two raters (>150 h; 10,194 walking bouts; gold standard).
For each device, commonly used analytical approaches were tested. Cadence (steps·min⁻¹) derived from detected steps was used to classify intensity (Tudor-Locke et al., 2019). Sensitivity, specificity, and weighted accuracy were computed for walking detection and intensity classification.
RESULTS:
Three main findings emerged:
(1) Aggregation ≥60 s overestimates walking duration and underestimates cadence, leading to an underestimation of moderate-to-vigorous PA in favor of light PA;
(2) With high temporal resolution (≤1 s), validity generally depends on sensor location, with thigh-, hip-, and ankle-worn devices better detecting short bouts; however, the wrist-worn StepCount-SSL algorithm (Small et al., 2024) achieved validity comparable to lower-limb and hip placements;
(3) At identical location and resolution, analytical choices (counts, vector magnitude, raw-signal-based or machine-learning methods) substantially affect validity.
CONCLUSION:
Free-living walking is highly fragmented. Ignoring short bouts lead to misestimating walking intensity, a key parameter of PA guidelines. In most cases, combining lower-limb or hip placement with high temporal resolution better reflects real-world walking dynamics. However, advanced wrist-based machine-learning approaches such as StepCount-SSL may overcome traditional wrist-placement limitations.
To accurately characterize intermittent walking in daily life, we recommend:
(1) high temporal resolution approaches (≤1 s or event-based detection), and
(2) sensor placement on the hip, thigh, or ankle, or the use of sufficiently validated wrist-based machine-learning algorithms (e.g., StepCount-SSL).